The Altair Community is migrating to a new platform to provide a better experience for you. In preparation for the migration, the Altair Community is on read-only mode from October 28 - November 6, 2024. Technical support via cases will continue to work as is. For any urgent requests from Students/Faculty members, please submit the form linked here
auto-find_precission/recall-breakeven-point
Contributor I
Hi there,
i am quite new to rapidminer and the whole subject of classification. my problem is in short the following:
any help would be appreciated.
thanks in advance
klaus
my process:
i am quite new to rapidminer and the whole subject of classification. my problem is in short the following:
- textclassification
- OHSUMED-91-dataset (~13 000 medical abstracts)
- i want to compare my results with the results of someone else who measured his classification with the precission-recall-breakeven-point
- metaCost-Operator
- thresholdFinder / thresholdApplier
any help would be appreciated.
thanks in advance
klaus
my process:
<?xml version="1.0" encoding="UTF-8" standalone="no"?>
<process version="5.3.015">
<context>
<input/>
<output/>
<macros/>
</context>
<operator activated="true" class="process" compatibility="5.3.015" expanded="true" name="Root">
<description><p>This process demonstrates how a threshold can be obtained from a soft classifier and applied to an independent test set.</p><ol><li>The learner used in this process makes soft predictions instead of crisp classifications. The prediction confidences delivered by all learners in RapidMiner which are able to handle nominal labels (classification) will be used as soft predictions. <br><icon>groups/24/learner</icon></li><li>The ThresholdFinder is used to determine the best threshold with respect to class weights. In this case, a wrong classification of the first class (negative) will cause costs five times bigger than the other error. <br><icon>groups/24/postprocessing</icon></li><li>Please note that a ModelApplier must be performed on the test set before a threshold can be found. Since this model must be applied again later, the model applier keeps the input model. <br><icon>operators/24/model_applier</icon></li><li>The IOConsumer ensures that the prediction is made on the correct data set. <br><icon>operators/24/io_consumer</icon></li><li>The last steps apply the model and the threshold on the data set at hand. <br><icon>groups/24/validation</icon></li></ol></description>
<parameter key="logverbosity" value="status"/>
<parameter key="random_seed" value="1903"/>
<process expanded="true">
<operator activated="true" class="text:process_document_from_file" compatibility="5.3.002" expanded="true" height="76" name="ProcessTrainingSet" width="90" x="45" y="30">
<list key="text_directories">
<parameter key="C23_less" value="C:\stuff\Dokumentenklassifikation\Datasets\Ohsumed_91\ohsumed-first-20000-divided-C23_notC23\training\C23_less"/>
<parameter key="notC23_less" value="C:\stuff\Dokumentenklassifikation\Datasets\Ohsumed_91\ohsumed-first-20000-divided-C23_notC23\training\notC23_less"/>
</list>
<process expanded="true">
<operator activated="true" class="text:tokenize" compatibility="5.3.002" expanded="true" height="60" name="Tokenize" width="90" x="45" y="30"/>
<operator activated="false" class="text:filter_by_length" compatibility="5.3.002" expanded="true" height="60" name="Filter Tokens (by Length)" width="90" x="179" y="30"/>
<operator activated="true" class="text:stem_snowball" compatibility="5.3.002" expanded="true" height="60" name="Stem (Snowball)" width="90" x="315" y="30"/>
<operator activated="true" class="text:transform_cases" compatibility="5.3.002" expanded="true" height="60" name="Transform Cases" width="90" x="623" y="30"/>
<connect from_port="document" to_op="Tokenize" to_port="document"/>
<connect from_op="Tokenize" from_port="document" to_op="Stem (Snowball)" to_port="document"/>
<connect from_op="Stem (Snowball)" from_port="document" to_op="Transform Cases" to_port="document"/>
<connect from_op="Transform Cases" from_port="document" to_port="document 1"/>
<portSpacing port="source_document" spacing="0"/>
<portSpacing port="sink_document 1" spacing="0"/>
<portSpacing port="sink_document 2" spacing="0"/>
</process>
</operator>
<operator activated="true" class="support_vector_machine_libsvm" compatibility="5.3.015" expanded="true" height="76" name="SVM (2)" width="90" x="246" y="30">
<parameter key="gamma" value="1.0"/>
<parameter key="C" value="1.0"/>
<parameter key="nu" value="0.4"/>
<list key="class_weights"/>
</operator>
<operator activated="true" class="text:process_document_from_file" compatibility="5.3.002" expanded="true" height="76" name="ProcessTestSet" width="90" x="112" y="165">
<list key="text_directories">
<parameter key="C23_less" value="C:\stuff\Dokumentenklassifikation\Datasets\Ohsumed_91\ohsumed-first-20000-divided-C23_notC23\test\C23_less"/>
<parameter key="notC23_less" value="C:\stuff\Dokumentenklassifikation\Datasets\Ohsumed_91\ohsumed-first-20000-divided-C23_notC23\test\notC23_less"/>
</list>
<process expanded="true">
<operator activated="true" class="text:tokenize" compatibility="5.3.002" expanded="true" height="60" name="Tokenize (2)" width="90" x="45" y="30"/>
<operator activated="false" class="text:filter_by_length" compatibility="5.3.002" expanded="true" height="60" name="Filter Tokens (2)" width="90" x="179" y="30"/>
<operator activated="true" class="text:stem_snowball" compatibility="5.3.002" expanded="true" height="60" name="Stem (2)" width="90" x="380" y="30"/>
<operator activated="true" class="text:transform_cases" compatibility="5.3.002" expanded="true" height="60" name="Transform Cases (2)" width="90" x="648" y="30"/>
<connect from_port="document" to_op="Tokenize (2)" to_port="document"/>
<connect from_op="Tokenize (2)" from_port="document" to_op="Stem (2)" to_port="document"/>
<connect from_op="Stem (2)" from_port="document" to_op="Transform Cases (2)" to_port="document"/>
<connect from_op="Transform Cases (2)" from_port="document" to_port="document 1"/>
<portSpacing port="source_document" spacing="0"/>
<portSpacing port="sink_document 1" spacing="0"/>
<portSpacing port="sink_document 2" spacing="0"/>
</process>
</operator>
<operator activated="true" class="apply_model" compatibility="5.3.015" expanded="true" height="76" name="TestModelApplier" width="90" x="380" y="120">
<list key="application_parameters"/>
</operator>
<operator activated="true" class="find_threshold" compatibility="5.3.015" expanded="true" height="76" name="ThresholdFinder" width="90" x="514" y="120">
<parameter key="misclassification_costs_first" value="1.335"/>
<parameter key="use_example_weights" value="false"/>
</operator>
<operator activated="true" class="apply_threshold" compatibility="5.3.015" expanded="true" height="76" name="ThresholdApplier" width="90" x="648" y="120"/>
<operator activated="true" class="performance_classification" compatibility="5.3.015" expanded="true" height="76" name="Performance (3)" width="90" x="782" y="120">
<parameter key="classification_error" value="true"/>
<parameter key="kappa" value="true"/>
<parameter key="weighted_mean_recall" value="true"/>
<parameter key="weighted_mean_precision" value="true"/>
<parameter key="spearman_rho" value="true"/>
<parameter key="kendall_tau" value="true"/>
<parameter key="absolute_error" value="true"/>
<parameter key="relative_error" value="true"/>
<parameter key="relative_error_lenient" value="true"/>
<parameter key="relative_error_strict" value="true"/>
<parameter key="normalized_absolute_error" value="true"/>
<parameter key="root_mean_squared_error" value="true"/>
<parameter key="root_relative_squared_error" value="true"/>
<parameter key="squared_error" value="true"/>
<parameter key="correlation" value="true"/>
<parameter key="squared_correlation" value="true"/>
<parameter key="cross-entropy" value="true"/>
<parameter key="margin" value="true"/>
<parameter key="soft_margin_loss" value="true"/>
<parameter key="logistic_loss" value="true"/>
<list key="class_weights"/>
</operator>
<connect from_op="ProcessTrainingSet" from_port="example set" to_op="SVM (2)" to_port="training set"/>
<connect from_op="ProcessTrainingSet" from_port="word list" to_op="ProcessTestSet" to_port="word list"/>
<connect from_op="SVM (2)" from_port="model" to_op="TestModelApplier" to_port="model"/>
<connect from_op="ProcessTestSet" from_port="example set" to_op="TestModelApplier" to_port="unlabelled data"/>
<connect from_op="TestModelApplier" from_port="labelled data" to_op="ThresholdFinder" to_port="example set"/>
<connect from_op="ThresholdFinder" from_port="example set" to_op="ThresholdApplier" to_port="example set"/>
<connect from_op="ThresholdFinder" from_port="threshold" to_op="ThresholdApplier" to_port="threshold"/>
<connect from_op="ThresholdApplier" from_port="example set" to_op="Performance (3)" to_port="labelled data"/>
<connect from_op="Performance (3)" from_port="performance" to_port="result 1"/>
<portSpacing port="source_input 1" spacing="0"/>
<portSpacing port="sink_result 1" spacing="36"/>
<portSpacing port="sink_result 2" spacing="0"/>
</process>
</operator>
</process>
0
Answers
to focus more on the threshold problem I built a new process with
sonar data and a Naive Bayes learner.
First of all: Do you want a lot of true positives (high precision)
or is it important to have less false positives (high recall) ?
You can not have both.
Especially if you are dealing with mines and rocks - like in the example process -
it is highly recommended for a submarine that _no_ mine is
predicted as a rock...
Please study the operator "Select Recall" used in the following process.
first of all, thank you much for providing me this example process. now with the "selectRecall"-operator i got another method for adjusting precission and recall. but still this is not exactly what i am looking for.
as mentioned above, i want to compare my classificator-results with those of some else (Thorsten Joachims: Text Categorization with Support Vector
Machines: Learning with Many Relevant Features).
and he meassured his classificator with the precission-recall-breakeven-point
the precission-recall-breakeven-point is defined as the point where precission and recall are equal.
i know that this is a very old and obsolete method for meassuring a classifier and that other values like the Foil’s information Gain, or the Likelihood-Ratio are much better for defining the quality of a classifier, which i will also use. but in order to be able to compare my results with those of Thorsten Joachims i need to find the precission-recall-breakeven-point. is there a way in rapidminer to automatically find the needed threshold or do i have to find it manually by trying different thresholds?
---
anther question: in your example when i type in "Mine" or "Rock" in the settings for "positive label" of the selectRecall-operator it seems that it has no effect. always "mine" is the positive label, even when i type in "Rock". also tried the remapBinominals-operator with no succes in changing the positive label: